Post on 10-Jul-2015
Helgi Páll Helgasonhelgi@perseptio.com
AGI 2012
Center for Analysis and Design
of Intelligent Agents,
Reykjavik University
Eric Niveleric.nivel@gmail.com
Kristinn R. Thórisson thorisson@gmail.com
Why is attention necessary for AGI?
What is constructivist methodology?
How to design an attention mechanism
AGI 2012
In the domain of intelligent systems, themanagement of system resources is typicallycalled “attention”
Biological (Human) Attention:• Selective concentration on particular aspects of the
environment while ignoring others
Artificial Attention:• Resource management and control mechanism to
assign limited system resources to processing of mostrelevant or important information
AGI 2012
AGI 2012
Time constraints
Abundant information Limited resources
ATTENTION(Resource management)
If we have detailed specifications of tasksand environments at design time, we alreadyknow:
• what kind of information is relevant to system operation
• how frequently the system has to sample information
• how quickly the system needs to make decisions
• the resource requirements of the system
AGI 2012
Major reduction in complexity (compared to real-world tasks and environments)
• Information filtering can be pre-programmed if characteristics of task-relevant information is known in advance
• Resource management and processing can be hand-tuned for specific tasks and environments in advance
Substantial dynamic adaption to tasks not required
AGI 2012
When tasks and environments are partially specified or unspecified at design time, the following is unknown:
• what kind of information is relevant to system operation
• how frequently the system has to sample information
• how quickly the system needs to make decisions
• the resource requirements of the system
AGI 2012
AGI 2012
Level o
f ab
stra
ctio
n
(sp
ecific
atio
n, g
oals
)
Operating environment
Narrow AI
AGI
Learning
AGI systems are not
supplied at design time
with sufficient explicit
initial knowledge to
achieve all goals
Must learn to realize high-
level goals in the
operating environment
Must learn to perceive and
act meaningfully in the
environment
Initial knowledge for lower
levels of abstraction is
incomplete
AGI system design must assume up-front:
• Incomplete knowledge of the world at boot time
• Real world complexity for environments and tasks
• All information is potentially important
• Not only limited, but insufficient resources at all times
• Dynamic tasks, environments and time constraints
AGI 2012
“Narrow” AI• Substantial dynamic adaptation to task not required
• Data filtering can be pre-programmed if characteristics of useful data known in advance
• Lower than real world task complexity Resource management and processing hand-tuned for specific scenarios
→ Attention not required (?)
AGI• Real world environmental complexity assumed up-front
• Computational resources for the AI assumed to be insufficient at all times Complexity calls for data filtering and intelligent resource allocation
• Environments and tasks unknown at implementation time Resource management must be adaptive
→ Demands strong focus on resource management and realtime processing
AGI 2012
A general attention mechanism for
implementation in AGI systems /
cognitive architectures
Replication of natural attention mechanisms is not a goal
(but work is biologically inspired at a high level)
AGI 2012
AGI 2012
Constructivist AI• “From Constructionist to Constructivist AI”, Thórisson 2009, BICA
proceedings
Targets systems that manage their own growth• From manually constructed initial state
(bootstrap/seed)
Methodology for building flexible AGI systems capable of autonomous self-reconfiguration at the architecture level
General• No limiting assumptions about tasks, environments or modalities• Architecture-independent
Adaptive• Learns from experience
Complete• Targets all operational information (internal and external)• Top-down + Bottom-up
Uniform• Data from all modalities treated identically (at cognitive levels of
processing)
AGI 2012
Attention functionality implemented in handful
of AGI systems
Limitations:
• Data-filtering only (control issues ignored)
• External information only (internal states ignored)
• Realtime processing not addressed
AGI 2012
Intellifest 2012
System-wide quantification of data relevance
Data relevance:• Goal-related (top-down)• Novelty / Unexpectedness (bottom-up)
System-wide quantification of process relevance
Process relevance:• Operational experience (“top-down”) Prior success or failure of individual processes to contribute to similar
or identical goals
• Available data (“bottom-up”) Available data may limit which processes can be run
Internal system: another dynamic and complexenvironment• Similar to the external task environment
Meta-cognitive functions responsible for system growthmust also process information selectively• Resources remain limited
Applying a single, unified attention mechanism to bothinternal and external environments significantlyfacilitates the creation of AGI systems capable ofperforming tasks and improving own performancewhile being subject to resource limitations and realtimeconstraints.
AGI 2012
Data-driven
Fine-grained
Predictive capabilities
Unified sensory processes
AGI 2012
Data item Process
Data relevance quantified in saliency parameter
Process relevance quantified in activationparameter
Execution Policy
Execute most active processes on most salientdata items
(data item must match process input specification)
The high-level role of attention is to quantify and assign saliency and activation values
Data items
Processes
New data
Sensory devices
Environment(Real world)
Actuation devices
Commands
Sampled data
Goals / Predictions
Attentional patterns
Derived
Matching
Data items
Processes
Data biasing
Top-downSampled data
Environment(Real world)
Sensory devices
Actuation devices
Commands
Data items
Processes
Bottom-up attentionalprocessess
Top-down
Bottom-up
Sampled data
Derived
Environment(Real world)
Sensory devices
Actuation devices
Goals / Predictions
Attentional patterns
Data biasing
CommandsEvaluation
Matching
Data items
Processes
Top-down
Bottom-up
Process biasing
Sampled data
Derived
Environment(Real world)
Sensory devices
Actuation devices
Bottom-up attentionalprocessess
Goals / Predictions
Attentional patterns
Data biasing
Commands
Data -> Process mapping
Evaluation
Matching
Data items
Processes
Top-down
Bottom-up
Contextualized process
performance history
Contextual process evaluation
Experience-based process activation
Sampled data
Derived
Data -> Process mapping
Environment(Real world)
Sensory devices
Actuation devices
Bottom-up attentionalprocessess
Evaluation
Goals / Predictions
Attentional patterns
Matching
Data biasing
Process biasing
Commands
Implementation of early version complete
Evaluation in progress
AGI 2012
Intellifest 2012
Work supported by the European Project HUMANOBS – Humanoids that Learn Socio-Comunnicative
Skills Through Observation (grant number 231453).
Intellifest 2012
Publications:
• Cognitive Architectures and Autonomy: A Comparative Review Kristinn R. Thórisson, Helgi Páll Helgason http://versita.metapress.com/content/052t1h656614848h/?p=4e1d01ba40e04d5d9f51da3977a8be04&pi=0
• Attention Capabilities for AI Systems Helgi Páll Helgason, Kristinn R. Thórisson
http://www.perseptio.com/publications/Helgason-ICINCO-2012.pdf
• On Attention Mechanisms for AGI Architectures: A Design Proposal (to be published) Helgi Páll Helgason, Kristinn R. Thórisson, Eric Nivel
http://www.perseptio.com/publications/Helgason-AGI-2012.pdf
AGI 2012
Thanks to:
Dr. Kristinn R. Thórisson
Eric Nivel
Kamilla Jóhannsdóttir